TL;DR
This paper introduces two novel methods for automatically adapting bilingual dictionaries to specific domains, crucial for improving machine translation and cross-language information retrieval. The proposed "Ratio of Associated Words" (RAW) method significantly outperforms a contextual similarity-based approach by leveraging word associations to identify domain-relevant translation equivalents, achieving higher accuracy and providing more interpretable results.
What Problem Does This Paper Address?
Bilingual dictionaries are essential tools for cross-lingual tasks, but the meaning and appropriate translation of words often change depending on the domain (e.g., "operation" in a military context versus a medical one). Traditional methods for adapting dictionaries to specific domains rely heavily on lexicographers' intuition, which is time-consuming and not scalable. While existing automated methods exist, they often require large parallel corpora (texts that are direct translations of each other), which are scarce. This paper addresses the need for automated domain adaptation methods that can work with "weakly comparable corpora" – collections of texts in two languages from the same domain but not direct translations.
Methods
The paper proposes and evaluates two methods for bilingual-dictionary adaptation:
- Contextual Similarity (CS) Method: This method assumes that a word and its translation equivalent appear in similar contexts. It characterizes a target word and its candidate translation equivalents using "context vectors" derived from associated words weighted by mutual information. The similarity between a candidate equivalent and the target word is then calculated based on the cosine similarity of their translated context vectors.
- Ratio of Associated Words (RAW) Method: This is the paper's novel contribution. It's based on the idea that words associated with a target word can suggest specific senses or translation equivalents. The method first extracts word associations from each language's corpus. Then, it aligns these associations translingually and iteratively calculates the correlation between translation equivalents and associated words. Finally, it selects translation equivalents based on the ratio of associated words that strongly suggest them. This method also includes a mechanism to identify "representative associated words" that help lexicographers validate the selected equivalents.
Datasets
- EDR Bilingual Dictionary: Used as the source bilingual dictionary.
- Wall Street Journal (WSJ) Corpus (English): A corpus of financial news articles (July 1994 - December 1995, 189MB).
- Nihon Keizai Shimbun (Nikkei) Corpus (Japanese): A corpus of Japanese economic news articles (December 1993 - November 1994, 275MB).
These corpora are described as "weakly comparable" because they are from the same domain (finance/business) but are not direct translations. The experiment focused on English nouns that appeared 10 or more times in the WSJ corpus as target words. For quantitative evaluation, 1,000 "pseudo target words" were created using randomly selected words from the WSJ corpus.
Key Insights and Ideas
- Domain-specific word meanings are crucial: The paper highlights that a word's translation can vary significantly across domains, making dictionary adaptation essential for accurate cross-lingual applications.
- Weakly comparable corpora are sufficient: The success of the RAW method demonstrates that direct parallel corpora are not always necessary for effective bilingual dictionary adaptation.
- Word associations as semantic clues: The RAW method's effectiveness stems from the insight that associated words can act as strong indicators of a target word's specific sense within a domain.
- Interpretability matters: The RAW method's ability to provide "representative associated words" alongside selected translation equivalents is a significant advantage, allowing for easier validation by human experts.
- Polysemy is a challenge for CS: The CS method struggles with polysemous words (words with multiple meanings) because their context vectors become a composite of different senses, diluting the similarity measure.
Key Results
- RAW outperforms CS: In a quantitative evaluation using pseudo target words, the RAW method achieved a maximum F-measure of 86%, while the CS method reached 82%.
- RAW is more interpretable: Qualitative evaluation showed that over 90% of the translation equivalents selected by the RAW method were judged as "definitely appropriate" by referring to the representative associated words.
- Computational feasibility: Both methods were found to be computationally feasible, with the initial word association extraction being the most time-consuming part, proportional to corpus size. The selection part was much faster for the RAW method compared to CS.
- RAW handles domain relevance better: Tables comparing the methods show that the RAW method often selects more domain-relevant translations, as evidenced by the provided representative associated words.
Limitations and Future Work
- Limitations of CS: The CS method struggles with polysemous words and is sensitive to the fixed number of translation equivalents it selects, making it difficult to adapt to varying domain specificity.
- Future Work: The authors plan to apply the RAW method to several other domains to further evaluate its effectiveness and its impact on application systems. They also suggest that the ratio of associated words can serve as a rough approximation of translation probability, potentially guiding more sophisticated word-sense disambiguation or translation selection.
Summary updated